Feature Extraction by Pairwise Discriminant Analysis

• Main Authors: Tzung-Ying Lin, 林宗穎
• Other Authors: Chin-Chun Chang
• Format: Others
• Language: en_US
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/10754768253685576274

碩士 === 國立臺灣海洋大學 === 資訊工程學系 === 94 === Abstract The linear discriminant analysis (LDA) is to find a linear transformation which can reduce the dimension of a feature vector and preserve most of the discriminant information of the vector. The LDA based on the Fisher criterion, known as the FDA, is the most popular approach. However, the FDA has some limitations. First, the reduced dimension is limited to the class number minus one. Besides, the LDA based on mixed class information, such as the FDA, may overemphasize or neglect the class distribution with magnitudes incompatible to the other classes. In this study, a novel scheme based on “pairwise discriminant analysis” is proposed to overcome the mentioned limitations of the FDA. The proposed method is a cascade of two different schemes, namely the FDA and the common-mean discriminant analysis (CMDA), and constitutes the desired linear transformation one column by one column. Based on the proposed measure to evaluate the effectiveness of a vector to discriminate the classes, the most effective one is selected at every iteration. The proposed scheme can also automatically determine the number of extracted features, which is a nontrivial problem for the LDA. The proposed approach was compared to various existing methods and the experimental results proved the feasibility of the proposed approach.